Dataset issues！

[tutu-star]

Hello, author. I want to know how to generate my own dataset in this format(instances_train2017_base.json, (instances_train2017_base_RN50relabel.json, instances_train2017_base_RN50x4relabel_pre.json) and instances_val2017_basetarget.json).

[dongwhfdyer]

I really want to know too!

[dongwhfdyer]

  if self.args.target_class_factor != 1.0 and not self.training:
                        if outputs_class.size(-1) == 66:
                            # COCO
                            target_index = [4, 5, 11, 12, 15, 16, 21, 23, 27, 29, 32, 34, 45, 47, 54, 58, 63]
                        elif outputs_class.size(-1) == 1204:
                            # LVIS
                            target_index = [12, 13, 16, 19, 20, 29, 30, 37, 38, 39, 41, 48, 50, 51, 62, 68, 70, 77, 81, 84, 92, 104, 105, 112, 116, 118, 122, 125, 129, 130, 135, 139, 141, 143, 146, 150, 154, 158, 160, 163, 166, 171, 178, 181, 195, 201, 208, 209, 213, 214, 221, 222, 230, 232, 233, 235, 236, 237, 239, 243, 244, 246, 249, 250, 256, 257, 261, 264, 265, 268, 269, 274, 280, 281, 286, 290, 291, 293, 294, 299, 300, 301, 303, 306, 309, 312, 315, 316, 320, 322, 325, 330, 332, 347, 348, 351, 352, 353, 354, 356, 361, 363, 364, 365, 367, 373, 375, 380, 381, 387, 388, 396, 397, 399, 404, 406, 409, 412, 413, 415, 419, 425, 426, 427, 430, 431, 434, 438, 445, 448, 455, 457, 466, 477, 478, 479, 480, 481, 485, 487, 490, 491, 502, 505, 507, 508, 512, 515, 517, 526, 531, 534, 537, 540, 541, 542, 544, 550, 556, 559, 560, 566, 567, 570, 571, 573, 574, 576, 579, 581, 582, 584, 593, 596, 598, 601, 602, 605, 609, 615, 617, 618, 619, 624, 631, 633, 634, 637, 639, 645, 647, 650, 656, 661, 662, 663, 664, 670, 671, 673, 677, 685, 687, 689, 690, 692, 701, 709, 711, 713, 721, 726, 728, 729, 732, 742, 751, 753, 754, 757, 758, 763, 768, 771, 777, 778, 782, 783, 784, 786, 787, 791, 795, 802, 804, 807, 808, 809, 811, 814, 819, 821, 822, 823, 828, 830, 848, 849, 850, 851, 852, 854, 855, 857, 858, 861, 863, 868, 872, 882, 885, 886, 889, 890, 891, 893, 901, 904, 907, 912, 913, 916, 917, 919, 924, 930, 936, 937, 938, 940, 941, 943, 944, 951, 955, 957, 968, 971, 973, 974, 982, 984, 986, 989, 990, 991, 993, 997, 1002, 1004, 1009, 1011, 1014, 1015, 1027, 1028, 1029, 1030, 1031, 1046, 1047, 1048, 1052, 1053, 1056, 1057, 1074, 1079, 1083, 1115, 1117, 1118, 1123, 1125, 1128, 1134, 1143, 1144, 1145, 1147, 1149, 1156, 1157, 1158, 1164, 1166, 1192]
                        else:
                            assert False, "the dataset may not be supported"

https://github.com/tgxs002/CORA/blob/c334a6d87bb19b23fa8c3374e7ff59213dc87e49/models/fast_detr.py#L240C13-L240C13
Hello, I have generated the json files for my own dataset. But I was stuck when it comes to evaluation.
what's the meaning of the target_index in this context? How to edit the target_index for new dataset?

[tutu-star]

  if self.args.target_class_factor != 1.0 and not self.training:
                        if outputs_class.size(-1) == 66:
                            # COCO
                            target_index = [4, 5, 11, 12, 15, 16, 21, 23, 27, 29, 32, 34, 45, 47, 54, 58, 63]
                        elif outputs_class.size(-1) == 1204:
                            # LVIS
                            target_index = [12, 13, 16, 19, 20, 29, 30, 37, 38, 39, 41, 48, 50, 51, 62, 68, 70, 77, 81, 84, 92, 104, 105, 112, 116, 118, 122, 125, 129, 130, 135, 139, 141, 143, 146, 150, 154, 158, 160, 163, 166, 171, 178, 181, 195, 201, 208, 209, 213, 214, 221, 222, 230, 232, 233, 235, 236, 237, 239, 243, 244, 246, 249, 250, 256, 257, 261, 264, 265, 268, 269, 274, 280, 281, 286, 290, 291, 293, 294, 299, 300, 301, 303, 306, 309, 312, 315, 316, 320, 322, 325, 330, 332, 347, 348, 351, 352, 353, 354, 356, 361, 363, 364, 365, 367, 373, 375, 380, 381, 387, 388, 396, 397, 399, 404, 406, 409, 412, 413, 415, 419, 425, 426, 427, 430, 431, 434, 438, 445, 448, 455, 457, 466, 477, 478, 479, 480, 481, 485, 487, 490, 491, 502, 505, 507, 508, 512, 515, 517, 526, 531, 534, 537, 540, 541, 542, 544, 550, 556, 559, 560, 566, 567, 570, 571, 573, 574, 576, 579, 581, 582, 584, 593, 596, 598, 601, 602, 605, 609, 615, 617, 618, 619, 624, 631, 633, 634, 637, 639, 645, 647, 650, 656, 661, 662, 663, 664, 670, 671, 673, 677, 685, 687, 689, 690, 692, 701, 709, 711, 713, 721, 726, 728, 729, 732, 742, 751, 753, 754, 757, 758, 763, 768, 771, 777, 778, 782, 783, 784, 786, 787, 791, 795, 802, 804, 807, 808, 809, 811, 814, 819, 821, 822, 823, 828, 830, 848, 849, 850, 851, 852, 854, 855, 857, 858, 861, 863, 868, 872, 882, 885, 886, 889, 890, 891, 893, 901, 904, 907, 912, 913, 916, 917, 919, 924, 930, 936, 937, 938, 940, 941, 943, 944, 951, 955, 957, 968, 971, 973, 974, 982, 984, 986, 989, 990, 991, 993, 997, 1002, 1004, 1009, 1011, 1014, 1015, 1027, 1028, 1029, 1030, 1031, 1046, 1047, 1048, 1052, 1053, 1056, 1057, 1074, 1079, 1083, 1115, 1117, 1118, 1123, 1125, 1128, 1134, 1143, 1144, 1145, 1147, 1149, 1156, 1157, 1158, 1164, 1166, 1192]
                        else:
                            assert False, "the dataset may not be supported"

https://github.com/tgxs002/CORA/blob/c334a6d87bb19b23fa8c3374e7ff59213dc87e49/models/fast_detr.py#L240C13-L240C13 Hello, I have generated the json files for my own dataset. But I was stuck when it comes to evaluation. what's the meaning of the target_index in this context? How to edit the target_index for new dataset?

Thank you. I think the corresponding one is the label index. Can you provide the code for generating JSON files? Thank you!

[dongwhfdyer]

I can only tell you what kind of json file would work. As you can see in these pictures below, I have only included the essential information in the json file. It's a dict with three field 'images', 'annotations', 'categories'.
About the 'categories' field, there's no need for 'embedding'.
The reason why I cannot provide the code is that I found I have this kind of format annotation file in the dataset I use. These are typical coco-format annotation. Someone has done it for me.
About the target_index, I still don't know how to use it.
Maybe we can have a discussion.

[tutu-star]

I can only tell you what kind of json file would work. As you can see in these pictures below, I have only included the essential information in the json file. It's a dict with three field 'images', 'annotations', 'categories'. About the 'categories' field, there's no need for 'embedding'. The reason why I cannot provide the code is that I found I have this kind of format annotation file in the dataset I use. These are typical coco-format annotation. Someone has done it for me. About the target_index, I still don't know how to use it. Maybe we can have a discussion.

Thank you very much for your reply and we look forward to your early completion. Thank you!

[RobertTang0]

Hello, I would like to ask how to use my own COCO type data set for training, which part of the code needs to be modified?

[fyang064]

I can only tell you what kind of json file would work. As you can see in these pictures below, I have only included the essential information in the json file. It's a dict with three field 'images', 'annotations', 'categories'. About the 'categories' field, there's no need for 'embedding'. The reason why I cannot provide the code is that I found I have this kind of format annotation file in the dataset I use. These are typical coco-format annotation. Someone has done it for me. About the target_index, I still don't know how to use it. Maybe we can have a discussion.

Hi Kuhn, your reply looks good. I'm just wondering if the categories in your json file are generated by the off-the-shelf language model since I didn't see anything like it in the json files provided by the authors.

[zifuwan]

I can only tell you what kind of json file would work. As you can see in these pictures below, I have only included the essential information in the json file. It's a dict with three field 'images', 'annotations', 'categories'. About the 'categories' field, there's no need for 'embedding'. The reason why I cannot provide the code is that I found I have this kind of format annotation file in the dataset I use. These are typical coco-format annotation. Someone has done it for me. About the target_index, I still don't know how to use it. Maybe we can have a discussion.

Hi, I'm also trying to reproduce the CLIP-Aligned Labeling. Regarding your said, "About the 'categories' field, there's no need for 'embedding'.", do you think we can use an un-trained region classifier to get a refined class score for each class? What I mean is, instead of getting the embeddings as you showed in the picture, shall we multiply the region embedding with text embedding to get a score?

[Luhui-Zhao]

Hello, I would like to ask how to use my own COCO type data set for training, which part of the code needs to be modified?

Hello, I also want to use this code to train my own data set, and there is also a problem like yours, I would like to communicate with you, may I ask if you have solved this problem? If it is convenient, you can add QQ, my QQ is 755476579, looking forward to your reply, thank you.

[tutu-star]

 Sorry,I can not get the text file.

…

------------------ 原始邮件 ------------------ 发件人: "tgxs002/CORA" ***@***.***>; 发送时间: 2024年4月13日(星期六) 上午10:39 ***@***.***>; ***@***.******@***.***>; 主题: Re: [tgxs002/CORA] Dataset issues！ (Issue #12) Hello, I would like to ask how to use my own COCO type data set for training, which part of the code needs to be modified? Hello, I also want to use this code to train my own data set, and there is also a problem like yours, I would like to communicate with you, may I ask if you have solved this problem? If it is convenient, you can add QQ, my QQ is 755476579, looking forward to your reply, thank you. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>